DFA-CON: A Contrastive Learning Approach for Detecting Copyright Infringement in DeepFake Art

TL;DR

DFA-CON targets copyright infringement in AI-generated art by learning a discriminative embedding space that pairs originals with their forged variants. It introduces forgery-aware supervised contrastive learning using a ResNet-50 encoder and a projection head to map images into a 2048-dimensional space (and 128-D for the projection), optimizing a SupCon objective to pull together original-forged pairs while separating unrelated images. Evaluated on the Deepfake Art Challenge dataset with attacks including Inpainting, Style Transfer, Adversarial Perturbation, and CutMix, DFA-CON substantially outperforms pretrained vision foundation models on overall metrics and most attack types, though CutMix remains challenging. The approach provides a practical, scalable pipeline for infringement verification and attribution in AI-generated art, with public release of code and checkpoints to foster further research in generative content forensics, and highlights the need for domain-specific methods in art forensics. Key mathematical components include the region-based infringement criterion $\| A(y)_\Omega - A(T(\hat{x}))_\Omega \| < f(|\Omega|) \cdot \delta$ and the SupCon loss $\mathcal{L}_i$ that leverages multiple positives per anchor, enabling robust, attribution-aware embeddings in the presence of diverse forgery types.

Abstract

Recent proliferation of generative AI tools for visual content creation-particularly in the context of visual artworks-has raised serious concerns about copyright infringement and forgery. The large-scale datasets used to train these models often contain a mixture of copyrighted and non-copyrighted artworks. Given the tendency of generative models to memorize training patterns, they are susceptible to varying degrees of copyright violation. Building on the recently proposed DeepfakeArt Challenge benchmark, this work introduces DFA-CON, a contrastive learning framework designed to detect copyright-infringing or forged AI-generated art. DFA-CON learns a discriminative representation space, posing affinity among original artworks and their forged counterparts within a contrastive learning framework. The model is trained across multiple attack types, including inpainting, style transfer, adversarial perturbation, and cutmix. Evaluation results demonstrate robust detection performance across most attack types, outperforming recent pretrained foundation models. Code and model checkpoints will be released publicly upon acceptance.

Figures (2)

• Figure 1: Methodology for training contrastive model to detect copyright infringements in DeepFake Art (DFA-CON)
• Figure 2: Copyright infringement detection pipeline